Water Bottle Defect Detection System Using Convolutional Neural Network

Water bottling is the major task in water and beverage production companies. One of the most concerning issues during the bottling process is to assure Encapsulation quality, such as no cover, high cover, crooked (curved) cover, this kind of issues could bring serious food safety problems. Therefore, it is necessary to detect these defects in the process of production and take out the unqualified bottles from production lines. In this study we presented the presented a mechanism to detect defects during bottling process using convolutional neural network algorithm. We collected water bottle images from different sources and categorized the collected images as the training, validation, and testing data, which contains 7800, 1800, and 184 images respectively. All categories have been labeled as: broken, normal water, normal cap, no cap, foreign object and scratched. After data organization, we have applied different image preprocessing and data augmentation. The processed data was used to train a developed convolutional neural network model. Additionally, you only look once version 3 (YOLO3) object detection algorithm was used to detect bottles from the input image. After all, the accuracy of the developed model is compared with the accuracy of the pre-trained VGG16 and ResNet50 model by training them on our dataset. In our study we are able to gate 96% of training accuracy and 99% of validation accuracy.